Forge Daily Dashboard — July 21, 2026

Lead Story

China Open Weights Just Took the Week: Kimi K3 and Qwen3.8 Outscale the US Frontier Within 24 Hours

Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter open-weight model claiming parity with GPT-5.6 Sol and Claude Fable 5. Hours later, Alibaba shipped Qwen3.8, a 2.4-trillion-parameter multimodal model ranked second only to Fable 5 internally. Two Chinese labs released a combined 5.2 trillion parameters of open weights in one day, each with a clear path to enterprise deployment. The open-vs-closed debate at the frontier is no longer theoretical.

Closed Western Frontier

GPT-5.6 Sol launched Thu Jul 17 at a three-tier price stack; now matched on benchmark claims by a Chinese open-weight model one day later.
Claude Fable 5 remains the internal benchmark leader, but its proprietary API is now the comparison point for open weights.
Thinking Machines Inkling shipped 975B parameters under Apache 2.0, adding a US open-weight counterpoint but at a smaller scale.

Open Weights + Chinese Labs

Kimi K3 — 2.8T parameters, Attention Residuals + Kimi Delta Attention, API live at $3/$15 per million tokens, full weights Jul 27.
Qwen3.8 — 2.4T parameters, fully multimodal (text / image / audio), planned open-weight release via Token Plan and Qoder.
Combined scale — 5.2T parameters released in one day by two Chinese labs, both with explicit open-weight roadmaps.

"The strategic unit is no longer the best API. It is the base model, the post-training platform, the inference ecosystem, and the auditable harness — together."

Model Surface

Capabilities, access, and pricing

Kimi K3: 2.8T total, 2.5x scaling efficiency from Attention Residuals, API live now, weights Jul 27.

Qwen3.8: 2.4T total, multimodal native, preview via Token Plan / Qoder, open-weight timeline TBD.

Inkling: 975B total / 41B active, 1M context, Apache 2.0, day-0 support in transformers, SGLang, llama.cpp.

Capital & Policy

Funding flows and macro signals

AI agent funding: $1.8B+ across 12+ deals in July, with European VCs narrowing bets as GenAI funding drops 50% in Q2.

Polymarket: 94-96% probability of a July 2026 Fed hold, with only 13-15% odds of any 2026 cut.

Portugal: €1.3B Horizon Europe funding captured, leading the EU Widening category.

Strategic Take

What changed for builders

The Chinese open-weight push is now a structural feature, not a one-off release. Enterprise buyers will compare proprietary API access against self-hosted 2T+ parameter models within months.

The durable advantage shifts to inference optimization, post-training platforms, and auditability — not model size alone.